Re: [bitcoin-dev] Should Graftroot be optional?

[Anthony Towns <aj@erisian.com.au>]

On Thu, May 31, 2018 at 05:25:04PM -0700, Pieter Wuille via bitcoin-dev wrote:
> The best argument for why Graftroot does not need to be optional I
> think was how Greg put it: "since the signer(s) could have signed an
> arbitrary transaction instead, being able to delegate is strictly less
> powerful.".

This seems persuasive to me. I think you could implement graftroot in
a way that makes this explicit:

 * A graftroot input has >=2 items on the witness stack, a signature,
   a script (S), and possibly witness elements for the script. The
   signature has a SIGHASH_GRAFTROOT bit set.

 * To validate the signature, a virtual transaction is constructed:

     nVersion = 1
     locktime = 0
     inputs = [(txhash, txoutidx, 0, "", 0xffffffff)]
     outputs = [(txvalue, len(S), S)]
     locktime = 0

   The signature is then checked against the virtual transaction.

 * If the signature is valid, the virtual transaction is discarded, and
   the script and witness elements are checked against the original tx.

I think this approach (or one like it) would make it clear that
graftroot is a simple optimisation, rather than changing the security
parameters. Some caveats:

 * You'd presumably want to disallow signatures with SIGHASH_GRAFTROOT
   from being used in signatures in scripts, so as not to end up having
   to support recursive graftroot.

 * Checking the script/witness against the original transaction instead
   of the virtual one cheats a bit, but something like it is necessary
   to ensure locktime/csv checks in the script S behave sanely. You
   could have the virtual transaction be treated as being confirmed in
   the same block as the original transaction instead though, I think.

 * You would need to use SIGHASH_NOINPUT (or similar) in conjuction
   to allow graftroot delegation prior to constructing the tx (otherwise
   the signature would be committing to txhash/txoutidx). BIP118 would
   still commit to txvalue, but would otherwise work fine, I think.

Cheers,
aj